Authors: Seth Goodman*, College of William and Mary, Daniel Runfola, College of William and Mary
Topics: Geographic Information Science and Systems, Spatial Analysis & Modeling
Keywords: convolutional neural networks, machine learning, gap filling
Session Type: Paper
Start / End Time: 5:00 PM / 6:40 PM
Room: Capitol Room, Omni, East
Presentation File: No File Uploaded
Recent work has proven the effectiveness of satellite imagery and convolutional neural network (CNN) based methods to predict poverty through human development indicators such as household consumption. A range of data sources, CNN architectures, and applications present challenges which are new both to the fields of geospatial data and machine learning. One of the most significant source of satellite imagery to date is the Landsat 7 ETM+ sensor, which offers global coverage from 1999 to the present. Although Landsat 7 has sufficient coverage to enable analysis of historic and contemporary conditions around the world, it has suffered from a scan line corrector (SLC) failure since 2003. This paper will explore the impact of satellite imagery data preparation, specifically gap filling methods for SLC-off Landsat 7 imagery, and CNN hyperparameters on the ability of CNNs to accurately classify nighttime light levels based on daytime scene samples. Gap filling methods to correct SLC-off images will be compared using hyperparameter optimization to identify effective sets of hyperparameters for each method. Based on the results of these tests, we present general guidelines for gap filling SLC-off Landsat 7 imagery, and starting points for identifying ideal hyperparameters for CNN based methods to predict human development indicators in future work.